Comment by AlecSchueler 6 hours ago
Your content moderation is quite hit and miss. I experimented with the question "how is babby formed?" and about 2/3 times it told me it couldn't respond to questions about sex, but then on the third attempt it would give a very explicit explanation.
“Works on my machine” actually isn’t a good enough response in this case, or to the comment about the video of the man being shot. LLMs are infamously easy to jailbreak and children are very good at getting around guardrails. You should at the very least be doing intense adversarial prompt testing but honestly this idea is just inherently poorly thought out. I guarantee you it’s going to expose children to harmful content
Thank you for pointing this out. We haven't been able to replicate this, but we will keep testing and work to improve on it.